Lead/Senior Research Scientist; Agentic RL
Job in
Saint Paul, Dakota County, Minnesota, 55123, USA
Listing for:
Thomson Reuters
Full Time
position
Listed on 2026-11-20
Job specializations:
-
Research/Development
Job Description & How to Apply Below
Location: Saint Paul
Job Description Thomson Reuters Labs builds and ships custom models for professional workflows requiring fiduciary-grade AI. The Thomson LLM family was developed by training leading open-weight models on decades of authoritative Thomson Reuters content, with thousands of hours of subject-matter-expert input. Owning the model layer is how we earn trust in the most regulated, highest-stakes professional workflows. See the Thomson 1.0 Technical Report on Hugging Face for more on the latest model release.
The research team trains and evaluates the Thomson models to perform at a frontier level for professional workflows. We work with a range of large open-weight models and use continual learning to improve them comprehensively, rather than pre-training from scratch or doing narrow fine-tuning. The training pipeline includes value alignment, data-centric mid-training on curated proprietary data, DPO, and reinforcement learning (RL) for product-ready tasks.
The evaluation group builds the measurement side: expert-annotated legal and tax benchmarks, agentic systems, and human preference studies with subject-matter experts. We regularly publish work that comes out of our training and evaluation efforts.
We are hiring scientists at Senior and Lead level across three areas of emphasis. In your application, please tell us which one or two fit you best:
Agentic reinforcement learning: post-training for agentic behaviour (planning, long-context reasoning, tool calls, long-horizon tasks inside real agent harnesses), multi-turn RL and policy optimisation, reward modelling and training-signal design.
Model training: continued pre-training and post-training of open-weight models, DPO and RL, training data and reward design, forgetting and replay in continued pre-training.
Model evaluation: benchmarks and targeted evaluations for complex legal and professional workflows, LLM-as-a-judge and rubric design, human preference studies with legal experts, and evaluation results that inform training objectives.
Key Responsibilities Design new benchmarks, training datasets, rewards and training signals to improve the Thomson model family, including for tasks where ground truth is scarce, delayed or partly subjective.
Design rigorous experiments and diagnostics (for example ablations) that separate algorithmic effects from data, reward and system artefacts, and carry results into production models.
Build and extend the evaluation suite to understand Thomson's strengths and weaknesses, and use the results to inform training objectives and technical trade-offs.
Work hand in hand with Research Engineers on our large-scale training and evaluation infrastructure, and with subject-matter experts in law, tax and other professional domains.
Coordinate with product teams and other model consumers to understand adoption requirements and drive integration.
Publish regularly at top-tier ML and NLP venues and represent Thomson Reuters Labs in the research community.
At Lead level: own a research agenda, turn emerging methods into a roadmap for our models, and mentor scientists and engineers.
Required Qualifications
PhD in machine learning, computer science or a related discipline, or equivalent practical experience.
Hands-on LLM post-training and/or evaluation experience, preferably at frontier model development scale, using Python and PyTorch.
A track record of research impact you can point to: publications at top-tier venues (for example NeurIPS, ICML, ICLR, ACL, EMNLP, COLM) and/or work that shaped an LLM development cycle.
Experience working closely with researchers and cross-functional stakeholders, including subject-matter experts, to turn ideas into practical training or evaluation assets.
Excellent written and oral communication,…
Position Requirements
10+ Years work experienceTo View & Apply for jobs on this site that accept applications from your location or country, tap the button below to make a Search.
(If this job is in fact in your jurisdiction, then you may be using a Proxy or VPN to access this site, and to progress further, you should change your connectivity to another mobile device or PC).
View / Apply for Jobs
Matching My Jurisdiction
(If this job is in fact in your jurisdiction, then you may be using a Proxy or VPN to access this site, and to progress further, you should change your connectivity to another mobile device or PC).
View / Apply for Jobs
Matching My Jurisdiction